4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
31 * This file is part of Lustre, http://www.lustre.org/
32 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Author: Isaac Huang <isaac@clusterfs.com>
39 #define DEBUG_SUBSYSTEM S_LNET
52 struct smoketest_rpc {
53 cfs_spinlock_t rpc_glock; /* global lock */
54 srpc_service_t *rpc_services[SRPC_SERVICE_MAX_ID + 1];
55 lnet_handle_eq_t rpc_lnet_eq; /* _the_ LNet event queue */
56 srpc_state_t rpc_state;
57 srpc_counters_t rpc_counters;
58 __u64 rpc_matchbits; /* matchbits counter */
62 int srpc_handle_rpc (swi_workitem_t *wi);
64 void srpc_get_counters (srpc_counters_t *cnt)
66 cfs_spin_lock(&srpc_data.rpc_glock);
67 *cnt = srpc_data.rpc_counters;
68 cfs_spin_unlock(&srpc_data.rpc_glock);
71 void srpc_set_counters (const srpc_counters_t *cnt)
73 cfs_spin_lock(&srpc_data.rpc_glock);
74 srpc_data.rpc_counters = *cnt;
75 cfs_spin_unlock(&srpc_data.rpc_glock);
79 srpc_add_bulk_page (srpc_bulk_t *bk, cfs_page_t *pg, int i)
81 LASSERT (i >= 0 && i < bk->bk_niov);
84 bk->bk_iovs[i].kiov_offset = 0;
85 bk->bk_iovs[i].kiov_page = pg;
86 bk->bk_iovs[i].kiov_len = CFS_PAGE_SIZE;
88 LASSERT (bk->bk_pages != NULL);
91 bk->bk_iovs[i].iov_len = CFS_PAGE_SIZE;
92 bk->bk_iovs[i].iov_base = cfs_page_address(pg);
98 srpc_free_bulk (srpc_bulk_t *bk)
103 LASSERT (bk != NULL);
105 LASSERT (bk->bk_pages != NULL);
108 for (i = 0; i < bk->bk_niov; i++) {
110 pg = bk->bk_iovs[i].kiov_page;
112 pg = bk->bk_pages[i];
114 if (pg == NULL) break;
120 LIBCFS_FREE(bk->bk_pages, sizeof(cfs_page_t *) * bk->bk_niov);
122 LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bk->bk_niov]));
127 srpc_alloc_bulk (int npages, int sink)
133 LASSERT (npages > 0 && npages <= LNET_MAX_IOV);
135 LIBCFS_ALLOC(bk, offsetof(srpc_bulk_t, bk_iovs[npages]));
137 CERROR ("Can't allocate descriptor for %d pages\n", npages);
141 memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[npages]));
143 bk->bk_niov = npages;
144 bk->bk_len = npages * CFS_PAGE_SIZE;
146 LIBCFS_ALLOC(pages, sizeof(cfs_page_t *) * npages);
148 LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[npages]));
149 CERROR ("Can't allocate page array for %d pages\n", npages);
153 memset(pages, 0, sizeof(cfs_page_t *) * npages);
154 bk->bk_pages = pages;
159 for (i = 0; i < npages; i++) {
160 cfs_page_t *pg = cfs_alloc_page(CFS_ALLOC_STD);
163 CERROR ("Can't allocate page %d of %d\n", i, npages);
168 srpc_add_bulk_page(bk, pg, i);
179 cfs_spin_lock(&srpc_data.rpc_glock);
180 id = srpc_data.rpc_matchbits++;
181 cfs_spin_unlock(&srpc_data.rpc_glock);
186 srpc_init_server_rpc (srpc_server_rpc_t *rpc,
187 srpc_service_t *sv, srpc_buffer_t *buffer)
189 memset(rpc, 0, sizeof(*rpc));
190 swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc,
191 sv->sv_id <= SRPC_FRAMEWORK_SERVICE_MAX_ID ?
192 CFS_WI_SCHED_SERIAL : CFS_WI_SCHED_ANY);
194 rpc->srpc_ev.ev_fired = 1; /* no event expected now */
196 rpc->srpc_service = sv;
197 rpc->srpc_reqstbuf = buffer;
198 rpc->srpc_peer = buffer->buf_peer;
199 rpc->srpc_self = buffer->buf_self;
200 LNetInvalidateHandle(&rpc->srpc_replymdh);
204 srpc_add_service (srpc_service_t *sv)
208 srpc_server_rpc_t *rpc;
210 LASSERT (sv->sv_concur > 0);
211 LASSERT (0 <= id && id <= SRPC_SERVICE_MAX_ID);
213 cfs_spin_lock(&srpc_data.rpc_glock);
215 LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
217 if (srpc_data.rpc_services[id] != NULL) {
218 cfs_spin_unlock(&srpc_data.rpc_glock);
222 srpc_data.rpc_services[id] = sv;
223 cfs_spin_unlock(&srpc_data.rpc_glock);
226 sv->sv_nposted_msg = 0;
227 sv->sv_shuttingdown = 0;
228 cfs_spin_lock_init(&sv->sv_lock);
229 CFS_INIT_LIST_HEAD(&sv->sv_free_rpcq);
230 CFS_INIT_LIST_HEAD(&sv->sv_active_rpcq);
231 CFS_INIT_LIST_HEAD(&sv->sv_posted_msgq);
232 CFS_INIT_LIST_HEAD(&sv->sv_blocked_msgq);
234 sv->sv_ev.ev_data = sv;
235 sv->sv_ev.ev_type = SRPC_REQUEST_RCVD;
237 for (i = 0; i < sv->sv_concur; i++) {
238 LIBCFS_ALLOC(rpc, sizeof(*rpc));
239 if (rpc == NULL) goto enomem;
241 cfs_list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
244 CDEBUG (D_NET, "Adding service: id %d, name %s, concurrency %d\n",
245 id, sv->sv_name, sv->sv_concur);
249 while (!cfs_list_empty(&sv->sv_free_rpcq)) {
250 rpc = cfs_list_entry(sv->sv_free_rpcq.next,
251 srpc_server_rpc_t, srpc_list);
252 cfs_list_del(&rpc->srpc_list);
253 LIBCFS_FREE(rpc, sizeof(*rpc));
256 cfs_spin_lock(&srpc_data.rpc_glock);
257 srpc_data.rpc_services[id] = NULL;
258 cfs_spin_unlock(&srpc_data.rpc_glock);
263 srpc_remove_service (srpc_service_t *sv)
267 cfs_spin_lock(&srpc_data.rpc_glock);
269 if (srpc_data.rpc_services[id] != sv) {
270 cfs_spin_unlock(&srpc_data.rpc_glock);
274 srpc_data.rpc_services[id] = NULL;
275 cfs_spin_unlock(&srpc_data.rpc_glock);
280 srpc_post_passive_rdma(int portal, __u64 matchbits, void *buf,
281 int len, int options, lnet_process_id_t peer,
282 lnet_handle_md_t *mdh, srpc_event_t *ev)
286 lnet_handle_me_t meh;
288 rc = LNetMEAttach(portal, peer, matchbits, 0,
289 LNET_UNLINK, LNET_INS_AFTER, &meh);
291 CERROR ("LNetMEAttach failed: %d\n", rc);
292 LASSERT (rc == -ENOMEM);
300 md.options = options;
301 md.eq_handle = srpc_data.rpc_lnet_eq;
303 rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
305 CERROR ("LNetMDAttach failed: %d\n", rc);
306 LASSERT (rc == -ENOMEM);
308 rc = LNetMEUnlink(meh);
314 "Posted passive RDMA: peer %s, portal %d, matchbits "LPX64"\n",
315 libcfs_id2str(peer), portal, matchbits);
320 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
321 int options, lnet_process_id_t peer, lnet_nid_t self,
322 lnet_handle_md_t *mdh, srpc_event_t *ev)
330 md.eq_handle = srpc_data.rpc_lnet_eq;
331 md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
332 md.options = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
334 rc = LNetMDBind(md, LNET_UNLINK, mdh);
336 CERROR ("LNetMDBind failed: %d\n", rc);
337 LASSERT (rc == -ENOMEM);
341 /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
342 * they're only meaningful for MDs attached to an ME (i.e. passive
344 if ((options & LNET_MD_OP_PUT) != 0) {
345 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
346 portal, matchbits, 0, 0);
348 LASSERT ((options & LNET_MD_OP_GET) != 0);
350 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
354 CERROR ("LNet%s(%s, %d, "LPD64") failed: %d\n",
355 ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
356 libcfs_id2str(peer), portal, matchbits, rc);
358 /* The forthcoming unlink event will complete this operation
359 * with failure, so fall through and return success here.
361 rc = LNetMDUnlink(*mdh);
365 "Posted active RDMA: peer %s, portal %u, matchbits "LPX64"\n",
366 libcfs_id2str(peer), portal, matchbits);
372 srpc_post_active_rqtbuf(lnet_process_id_t peer, int service, void *buf,
373 int len, lnet_handle_md_t *mdh, srpc_event_t *ev)
378 if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
379 portal = SRPC_REQUEST_PORTAL;
381 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
383 rc = srpc_post_active_rdma(portal, service, buf, len,
384 LNET_MD_OP_PUT, peer,
385 LNET_NID_ANY, mdh, ev);
390 srpc_post_passive_rqtbuf(int service, void *buf, int len,
391 lnet_handle_md_t *mdh, srpc_event_t *ev)
395 lnet_process_id_t any = {0};
397 any.nid = LNET_NID_ANY;
398 any.pid = LNET_PID_ANY;
400 if (service > SRPC_FRAMEWORK_SERVICE_MAX_ID)
401 portal = SRPC_REQUEST_PORTAL;
403 portal = SRPC_FRAMEWORK_REQUEST_PORTAL;
405 rc = srpc_post_passive_rdma(portal, service, buf, len,
406 LNET_MD_OP_PUT, any, mdh, ev);
411 srpc_service_post_buffer (srpc_service_t *sv, srpc_buffer_t *buf)
413 srpc_msg_t *msg = &buf->buf_msg;
416 LASSERT (!sv->sv_shuttingdown);
418 LNetInvalidateHandle(&buf->buf_mdh);
419 cfs_list_add(&buf->buf_list, &sv->sv_posted_msgq);
420 sv->sv_nposted_msg++;
421 cfs_spin_unlock(&sv->sv_lock);
423 rc = srpc_post_passive_rqtbuf(sv->sv_id, msg, sizeof(*msg),
424 &buf->buf_mdh, &sv->sv_ev);
426 /* At this point, a RPC (new or delayed) may have arrived in
427 * msg and its event handler has been called. So we must add
428 * buf to sv_posted_msgq _before_ dropping sv_lock */
430 cfs_spin_lock(&sv->sv_lock);
433 if (sv->sv_shuttingdown) {
434 cfs_spin_unlock(&sv->sv_lock);
436 /* srpc_shutdown_service might have tried to unlink me
437 * when my buf_mdh was still invalid */
438 LNetMDUnlink(buf->buf_mdh);
440 cfs_spin_lock(&sv->sv_lock);
445 sv->sv_nposted_msg--;
446 if (sv->sv_shuttingdown) return rc;
448 cfs_list_del(&buf->buf_list);
450 cfs_spin_unlock(&sv->sv_lock);
451 LIBCFS_FREE(buf, sizeof(*buf));
452 cfs_spin_lock(&sv->sv_lock);
457 srpc_service_add_buffers (srpc_service_t *sv, int nbuffer)
463 LASSERTF (nbuffer > 0,
464 "nbuffer must be positive: %d\n", nbuffer);
466 for (posted = 0; posted < nbuffer; posted++) {
467 LIBCFS_ALLOC(buf, sizeof(*buf));
468 if (buf == NULL) break;
470 cfs_spin_lock(&sv->sv_lock);
471 rc = srpc_service_post_buffer(sv, buf);
472 cfs_spin_unlock(&sv->sv_lock);
481 srpc_service_remove_buffers (srpc_service_t *sv, int nbuffer)
483 LASSERTF (nbuffer > 0,
484 "nbuffer must be positive: %d\n", nbuffer);
486 cfs_spin_lock(&sv->sv_lock);
488 LASSERT (sv->sv_nprune >= 0);
489 LASSERT (!sv->sv_shuttingdown);
491 sv->sv_nprune += nbuffer;
493 cfs_spin_unlock(&sv->sv_lock);
497 /* returns 1 if sv has finished, otherwise 0 */
499 srpc_finish_service (srpc_service_t *sv)
501 srpc_server_rpc_t *rpc;
504 cfs_spin_lock(&sv->sv_lock);
506 LASSERT (sv->sv_shuttingdown); /* srpc_shutdown_service called */
508 if (sv->sv_nposted_msg != 0 || !cfs_list_empty(&sv->sv_active_rpcq)) {
510 "waiting for %d posted buffers to unlink and "
511 "in-flight RPCs to die.\n",
514 if (!cfs_list_empty(&sv->sv_active_rpcq)) {
515 rpc = cfs_list_entry(sv->sv_active_rpcq.next,
516 srpc_server_rpc_t, srpc_list);
517 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, "
518 "wi %s scheduled %d running %d, "
519 "ev fired %d type %d status %d lnet %d\n",
520 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
521 swi_state2str(rpc->srpc_wi.swi_state),
522 rpc->srpc_wi.swi_workitem.wi_scheduled,
523 rpc->srpc_wi.swi_workitem.wi_running,
524 rpc->srpc_ev.ev_fired,
525 rpc->srpc_ev.ev_type,
526 rpc->srpc_ev.ev_status,
527 rpc->srpc_ev.ev_lnet);
530 cfs_spin_unlock(&sv->sv_lock);
534 cfs_spin_unlock(&sv->sv_lock); /* no lock needed from now on */
539 if (!cfs_list_empty(&sv->sv_posted_msgq))
540 q = &sv->sv_posted_msgq;
541 else if (!cfs_list_empty(&sv->sv_blocked_msgq))
542 q = &sv->sv_blocked_msgq;
546 buf = cfs_list_entry(q->next, srpc_buffer_t, buf_list);
547 cfs_list_del(&buf->buf_list);
549 LIBCFS_FREE(buf, sizeof(*buf));
552 while (!cfs_list_empty(&sv->sv_free_rpcq)) {
553 rpc = cfs_list_entry(sv->sv_free_rpcq.next,
554 srpc_server_rpc_t, srpc_list);
555 cfs_list_del(&rpc->srpc_list);
556 LIBCFS_FREE(rpc, sizeof(*rpc));
562 /* called with sv->sv_lock held */
564 srpc_service_recycle_buffer (srpc_service_t *sv, srpc_buffer_t *buf)
566 if (sv->sv_shuttingdown) goto free;
568 if (sv->sv_nprune == 0) {
569 if (srpc_service_post_buffer(sv, buf) != 0)
570 CWARN ("Failed to post %s buffer\n", sv->sv_name);
576 cfs_spin_unlock(&sv->sv_lock);
577 LIBCFS_FREE(buf, sizeof(*buf));
578 cfs_spin_lock(&sv->sv_lock);
581 /* called with srpc_service_t::sv_lock held */
583 srpc_schedule_server_rpc (srpc_server_rpc_t *rpc)
585 swi_schedule_workitem(&rpc->srpc_wi);
589 srpc_abort_service (srpc_service_t *sv)
591 srpc_server_rpc_t *rpc;
593 cfs_spin_lock(&sv->sv_lock);
595 CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
596 sv->sv_id, sv->sv_name);
598 /* schedule in-flight RPCs to notice the abort, NB:
599 * racing with incoming RPCs; complete fix should make test
600 * RPCs carry session ID in its headers */
601 cfs_list_for_each_entry (rpc, &sv->sv_active_rpcq, srpc_list) {
602 rpc->srpc_aborted = 1;
603 srpc_schedule_server_rpc(rpc);
606 cfs_spin_unlock(&sv->sv_lock);
611 srpc_shutdown_service (srpc_service_t *sv)
613 srpc_server_rpc_t *rpc;
616 cfs_spin_lock(&sv->sv_lock);
618 CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
619 sv->sv_id, sv->sv_name);
621 sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
623 /* schedule in-flight RPCs to notice the shutdown */
624 cfs_list_for_each_entry_typed (rpc, &sv->sv_active_rpcq,
625 srpc_server_rpc_t, srpc_list) {
626 srpc_schedule_server_rpc(rpc);
629 cfs_spin_unlock(&sv->sv_lock);
631 /* OK to traverse sv_posted_msgq without lock, since no one
632 * touches sv_posted_msgq now */
633 cfs_list_for_each_entry_typed (buf, &sv->sv_posted_msgq,
634 srpc_buffer_t, buf_list)
635 LNetMDUnlink(buf->buf_mdh);
641 srpc_send_request (srpc_client_rpc_t *rpc)
643 srpc_event_t *ev = &rpc->crpc_reqstev;
648 ev->ev_type = SRPC_REQUEST_SENT;
650 rc = srpc_post_active_rqtbuf(rpc->crpc_dest, rpc->crpc_service,
651 &rpc->crpc_reqstmsg, sizeof(srpc_msg_t),
652 &rpc->crpc_reqstmdh, ev);
654 LASSERT (rc == -ENOMEM);
655 ev->ev_fired = 1; /* no more event expected */
661 srpc_prepare_reply (srpc_client_rpc_t *rpc)
663 srpc_event_t *ev = &rpc->crpc_replyev;
664 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
669 ev->ev_type = SRPC_REPLY_RCVD;
671 *id = srpc_next_id();
673 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
674 &rpc->crpc_replymsg, sizeof(srpc_msg_t),
675 LNET_MD_OP_PUT, rpc->crpc_dest,
676 &rpc->crpc_replymdh, ev);
678 LASSERT (rc == -ENOMEM);
679 ev->ev_fired = 1; /* no more event expected */
685 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
687 srpc_bulk_t *bk = &rpc->crpc_bulk;
688 srpc_event_t *ev = &rpc->crpc_bulkev;
689 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
693 LASSERT (bk->bk_niov <= LNET_MAX_IOV);
695 if (bk->bk_niov == 0) return 0; /* nothing to do */
697 opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
701 opt |= LNET_MD_IOVEC;
706 ev->ev_type = SRPC_BULK_REQ_RCVD;
708 *id = srpc_next_id();
710 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, *id,
711 &bk->bk_iovs[0], bk->bk_niov, opt,
712 rpc->crpc_dest, &bk->bk_mdh, ev);
714 LASSERT (rc == -ENOMEM);
715 ev->ev_fired = 1; /* no more event expected */
721 srpc_do_bulk (srpc_server_rpc_t *rpc)
723 srpc_event_t *ev = &rpc->srpc_ev;
724 srpc_bulk_t *bk = rpc->srpc_bulk;
725 __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
729 LASSERT (bk != NULL);
731 opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
735 opt |= LNET_MD_IOVEC;
740 ev->ev_type = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
742 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
743 &bk->bk_iovs[0], bk->bk_niov, opt,
744 rpc->srpc_peer, rpc->srpc_self,
747 ev->ev_fired = 1; /* no more event expected */
751 /* only called from srpc_handle_rpc */
753 srpc_server_rpc_done (srpc_server_rpc_t *rpc, int status)
755 srpc_service_t *sv = rpc->srpc_service;
756 srpc_buffer_t *buffer;
758 LASSERT (status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
760 rpc->srpc_status = status;
762 CDEBUG_LIMIT (status == 0 ? D_NET : D_NETERROR,
763 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
764 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
765 swi_state2str(rpc->srpc_wi.swi_state), status);
768 cfs_spin_lock(&srpc_data.rpc_glock);
769 srpc_data.rpc_counters.rpcs_dropped++;
770 cfs_spin_unlock(&srpc_data.rpc_glock);
773 if (rpc->srpc_done != NULL)
774 (*rpc->srpc_done) (rpc);
775 LASSERT (rpc->srpc_bulk == NULL);
777 cfs_spin_lock(&sv->sv_lock);
779 if (rpc->srpc_reqstbuf != NULL) {
780 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
781 * sv won't go away for sv_active_rpcq must not be empty */
782 srpc_service_recycle_buffer(sv, rpc->srpc_reqstbuf);
783 rpc->srpc_reqstbuf = NULL;
786 cfs_list_del(&rpc->srpc_list); /* from sv->sv_active_rpcq */
789 * No one can schedule me now since:
790 * - I'm not on sv_active_rpcq.
791 * - all LNet events have been fired.
792 * Cancel pending schedules and prevent future schedule attempts:
794 LASSERT (rpc->srpc_ev.ev_fired);
795 swi_kill_workitem(&rpc->srpc_wi);
797 if (!sv->sv_shuttingdown && !cfs_list_empty(&sv->sv_blocked_msgq)) {
798 buffer = cfs_list_entry(sv->sv_blocked_msgq.next,
799 srpc_buffer_t, buf_list);
800 cfs_list_del(&buffer->buf_list);
802 srpc_init_server_rpc(rpc, sv, buffer);
803 cfs_list_add_tail(&rpc->srpc_list, &sv->sv_active_rpcq);
804 srpc_schedule_server_rpc(rpc);
806 cfs_list_add(&rpc->srpc_list, &sv->sv_free_rpcq);
809 cfs_spin_unlock(&sv->sv_lock);
813 /* handles an incoming RPC */
815 srpc_handle_rpc (swi_workitem_t *wi)
817 srpc_server_rpc_t *rpc = wi->swi_workitem.wi_data;
818 srpc_service_t *sv = rpc->srpc_service;
819 srpc_event_t *ev = &rpc->srpc_ev;
822 LASSERT (wi == &rpc->srpc_wi);
824 cfs_spin_lock(&sv->sv_lock);
826 if (sv->sv_shuttingdown || rpc->srpc_aborted) {
827 cfs_spin_unlock(&sv->sv_lock);
829 if (rpc->srpc_bulk != NULL)
830 LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
831 LNetMDUnlink(rpc->srpc_replymdh);
833 if (ev->ev_fired) { /* no more event, OK to finish */
834 srpc_server_rpc_done(rpc, -ESHUTDOWN);
840 cfs_spin_unlock(&sv->sv_lock);
842 switch (wi->swi_state) {
845 case SWI_STATE_NEWBORN: {
847 srpc_generic_reply_t *reply;
849 msg = &rpc->srpc_reqstbuf->buf_msg;
850 reply = &rpc->srpc_replymsg.msg_body.reply;
852 if (msg->msg_magic == 0) {
853 /* moaned already in srpc_lnet_ev_handler */
855 } else if (msg->msg_version != SRPC_MSG_VERSION &&
856 msg->msg_version != __swab32(SRPC_MSG_VERSION)) {
857 CWARN ("Version mismatch: %u, %u expected, from %s\n",
858 msg->msg_version, SRPC_MSG_VERSION,
859 libcfs_id2str(rpc->srpc_peer));
860 reply->status = EPROTO;
863 rc = (*sv->sv_handler) (rpc);
864 LASSERT (reply->status == 0 || !rpc->srpc_bulk);
868 srpc_server_rpc_done(rpc, rc);
872 wi->swi_state = SWI_STATE_BULK_STARTED;
874 if (rpc->srpc_bulk != NULL) {
875 rc = srpc_do_bulk(rpc);
877 return 0; /* wait for bulk */
879 LASSERT (ev->ev_fired);
883 case SWI_STATE_BULK_STARTED:
884 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
886 if (rpc->srpc_bulk != NULL) {
889 if (sv->sv_bulk_ready != NULL)
890 rc = (*sv->sv_bulk_ready) (rpc, rc);
893 srpc_server_rpc_done(rpc, rc);
898 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
899 rc = srpc_send_reply(rpc);
901 return 0; /* wait for reply */
902 srpc_server_rpc_done(rpc, rc);
905 case SWI_STATE_REPLY_SUBMITTED:
907 CERROR("RPC %p: bulk %p, service %d\n",
908 rpc, rpc->srpc_bulk, rpc->srpc_service->sv_id);
909 CERROR("Event: status %d, type %d, lnet %d\n",
910 ev->ev_status, ev->ev_type, ev->ev_lnet);
911 LASSERT (ev->ev_fired);
914 wi->swi_state = SWI_STATE_DONE;
915 srpc_server_rpc_done(rpc, ev->ev_status);
923 srpc_client_rpc_expired (void *data)
925 srpc_client_rpc_t *rpc = data;
927 CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
928 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
931 cfs_spin_lock(&rpc->crpc_lock);
933 rpc->crpc_timeout = 0;
934 srpc_abort_rpc(rpc, -ETIMEDOUT);
936 cfs_spin_unlock(&rpc->crpc_lock);
938 cfs_spin_lock(&srpc_data.rpc_glock);
939 srpc_data.rpc_counters.rpcs_expired++;
940 cfs_spin_unlock(&srpc_data.rpc_glock);
945 srpc_add_client_rpc_timer (srpc_client_rpc_t *rpc)
947 stt_timer_t *timer = &rpc->crpc_timer;
949 if (rpc->crpc_timeout == 0) return;
951 CFS_INIT_LIST_HEAD(&timer->stt_list);
952 timer->stt_data = rpc;
953 timer->stt_func = srpc_client_rpc_expired;
954 timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
955 cfs_time_current_sec());
956 stt_add_timer(timer);
961 * Called with rpc->crpc_lock held.
963 * Upon exit the RPC expiry timer is not queued and the handler is not
964 * running on any CPU. */
966 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
968 /* timer not planted or already exploded */
969 if (rpc->crpc_timeout == 0) return;
971 /* timer sucessfully defused */
972 if (stt_del_timer(&rpc->crpc_timer)) return;
975 /* timer detonated, wait for it to explode */
976 while (rpc->crpc_timeout != 0) {
977 cfs_spin_unlock(&rpc->crpc_lock);
981 cfs_spin_lock(&rpc->crpc_lock);
984 LBUG(); /* impossible in single-threaded runtime */
990 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
992 swi_workitem_t *wi = &rpc->crpc_wi;
994 LASSERT (status != 0 || wi->swi_state == SWI_STATE_DONE);
996 cfs_spin_lock(&rpc->crpc_lock);
998 rpc->crpc_closed = 1;
999 if (rpc->crpc_status == 0)
1000 rpc->crpc_status = status;
1002 srpc_del_client_rpc_timer(rpc);
1004 CDEBUG_LIMIT ((status == 0) ? D_NET : D_NETERROR,
1005 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1006 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1007 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1010 * No one can schedule me now since:
1011 * - RPC timer has been defused.
1012 * - all LNet events have been fired.
1013 * - crpc_closed has been set, preventing srpc_abort_rpc from
1015 * Cancel pending schedules and prevent future schedule attempts:
1017 LASSERT (!srpc_event_pending(rpc));
1018 swi_kill_workitem(wi);
1020 cfs_spin_unlock(&rpc->crpc_lock);
1022 (*rpc->crpc_done) (rpc);
1026 /* sends an outgoing RPC */
1028 srpc_send_rpc (swi_workitem_t *wi)
1031 srpc_client_rpc_t *rpc = wi->swi_workitem.wi_data;
1032 srpc_msg_t *reply = &rpc->crpc_replymsg;
1033 int do_bulk = rpc->crpc_bulk.bk_niov > 0;
1035 LASSERT (rpc != NULL);
1036 LASSERT (wi == &rpc->crpc_wi);
1038 cfs_spin_lock(&rpc->crpc_lock);
1040 if (rpc->crpc_aborted) {
1041 cfs_spin_unlock(&rpc->crpc_lock);
1045 cfs_spin_unlock(&rpc->crpc_lock);
1047 switch (wi->swi_state) {
1050 case SWI_STATE_NEWBORN:
1051 LASSERT (!srpc_event_pending(rpc));
1053 rc = srpc_prepare_reply(rpc);
1055 srpc_client_rpc_done(rpc, rc);
1059 rc = srpc_prepare_bulk(rpc);
1062 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1063 rc = srpc_send_request(rpc);
1066 case SWI_STATE_REQUEST_SUBMITTED:
1067 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1068 * order; however, they're processed in a strict order:
1069 * rqt, rpy, and bulk. */
1070 if (!rpc->crpc_reqstev.ev_fired) break;
1072 rc = rpc->crpc_reqstev.ev_status;
1075 wi->swi_state = SWI_STATE_REQUEST_SENT;
1076 /* perhaps more events, fall thru */
1077 case SWI_STATE_REQUEST_SENT: {
1078 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1080 if (!rpc->crpc_replyev.ev_fired) break;
1082 rc = rpc->crpc_replyev.ev_status;
1085 if ((reply->msg_type != type &&
1086 reply->msg_type != __swab32(type)) ||
1087 (reply->msg_magic != SRPC_MSG_MAGIC &&
1088 reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1089 CWARN ("Bad message from %s: type %u (%d expected),"
1090 " magic %u (%d expected).\n",
1091 libcfs_id2str(rpc->crpc_dest),
1092 reply->msg_type, type,
1093 reply->msg_magic, SRPC_MSG_MAGIC);
1098 if (do_bulk && reply->msg_body.reply.status != 0) {
1099 CWARN ("Remote error %d at %s, unlink bulk buffer in "
1100 "case peer didn't initiate bulk transfer\n",
1101 reply->msg_body.reply.status,
1102 libcfs_id2str(rpc->crpc_dest));
1103 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1106 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1108 case SWI_STATE_REPLY_RECEIVED:
1109 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1111 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1113 /* Bulk buffer was unlinked due to remote error. Clear error
1114 * since reply buffer still contains valid data.
1115 * NB rpc->crpc_done shouldn't look into bulk data in case of
1117 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1118 rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1121 wi->swi_state = SWI_STATE_DONE;
1122 srpc_client_rpc_done(rpc, rc);
1127 cfs_spin_lock(&rpc->crpc_lock);
1128 srpc_abort_rpc(rpc, rc);
1129 cfs_spin_unlock(&rpc->crpc_lock);
1133 if (rpc->crpc_aborted) {
1134 LNetMDUnlink(rpc->crpc_reqstmdh);
1135 LNetMDUnlink(rpc->crpc_replymdh);
1136 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1138 if (!srpc_event_pending(rpc)) {
1139 srpc_client_rpc_done(rpc, -EINTR);
1147 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1148 int nbulkiov, int bulklen,
1149 void (*rpc_done)(srpc_client_rpc_t *),
1150 void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1152 srpc_client_rpc_t *rpc;
1154 LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1155 crpc_bulk.bk_iovs[nbulkiov]));
1159 srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1160 bulklen, rpc_done, rpc_fini, priv);
1164 /* called with rpc->crpc_lock held */
1166 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1170 if (rpc->crpc_aborted || /* already aborted */
1171 rpc->crpc_closed) /* callback imminent */
1175 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1176 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1177 swi_state2str(rpc->crpc_wi.swi_state), why);
1179 rpc->crpc_aborted = 1;
1180 rpc->crpc_status = why;
1181 swi_schedule_workitem(&rpc->crpc_wi);
1185 /* called with rpc->crpc_lock held */
1187 srpc_post_rpc (srpc_client_rpc_t *rpc)
1189 LASSERT (!rpc->crpc_aborted);
1190 LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1191 LASSERT ((rpc->crpc_bulk.bk_len & ~CFS_PAGE_MASK) == 0);
1193 CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1194 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1197 srpc_add_client_rpc_timer(rpc);
1198 swi_schedule_workitem(&rpc->crpc_wi);
1204 srpc_send_reply (srpc_server_rpc_t *rpc)
1206 srpc_event_t *ev = &rpc->srpc_ev;
1207 srpc_msg_t *msg = &rpc->srpc_replymsg;
1208 srpc_buffer_t *buffer = rpc->srpc_reqstbuf;
1209 srpc_service_t *sv = rpc->srpc_service;
1213 LASSERT (buffer != NULL);
1214 rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1216 cfs_spin_lock(&sv->sv_lock);
1218 if (!sv->sv_shuttingdown &&
1219 sv->sv_id > SRPC_FRAMEWORK_SERVICE_MAX_ID) {
1220 /* Repost buffer before replying since test client
1221 * might send me another RPC once it gets the reply */
1222 if (srpc_service_post_buffer(sv, buffer) != 0)
1223 CWARN ("Failed to repost %s buffer\n", sv->sv_name);
1224 rpc->srpc_reqstbuf = NULL;
1227 cfs_spin_unlock(&sv->sv_lock);
1231 ev->ev_type = SRPC_REPLY_SENT;
1233 msg->msg_magic = SRPC_MSG_MAGIC;
1234 msg->msg_version = SRPC_MSG_VERSION;
1235 msg->msg_type = srpc_service2reply(sv->sv_id);
1237 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1238 sizeof(*msg), LNET_MD_OP_PUT,
1239 rpc->srpc_peer, rpc->srpc_self,
1240 &rpc->srpc_replymdh, ev);
1242 ev->ev_fired = 1; /* no more event expected */
1246 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1248 srpc_lnet_ev_handler (lnet_event_t *ev)
1250 srpc_event_t *rpcev = ev->md.user_ptr;
1251 srpc_client_rpc_t *crpc;
1252 srpc_server_rpc_t *srpc;
1253 srpc_buffer_t *buffer;
1256 srpc_msg_type_t type;
1258 LASSERT (!cfs_in_interrupt());
1260 if (ev->status != 0) {
1261 cfs_spin_lock(&srpc_data.rpc_glock);
1262 srpc_data.rpc_counters.errors++;
1263 cfs_spin_unlock(&srpc_data.rpc_glock);
1266 rpcev->ev_lnet = ev->type;
1268 switch (rpcev->ev_type) {
1270 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1271 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1273 case SRPC_REQUEST_SENT:
1274 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1275 cfs_spin_lock(&srpc_data.rpc_glock);
1276 srpc_data.rpc_counters.rpcs_sent++;
1277 cfs_spin_unlock(&srpc_data.rpc_glock);
1279 case SRPC_REPLY_RCVD:
1280 case SRPC_BULK_REQ_RCVD:
1281 crpc = rpcev->ev_data;
1283 if (rpcev != &crpc->crpc_reqstev &&
1284 rpcev != &crpc->crpc_replyev &&
1285 rpcev != &crpc->crpc_bulkev) {
1286 CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1287 rpcev, crpc, &crpc->crpc_reqstev,
1288 &crpc->crpc_replyev, &crpc->crpc_bulkev);
1289 CERROR("Bad event: status %d, type %d, lnet %d\n",
1290 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1294 cfs_spin_lock(&crpc->crpc_lock);
1296 LASSERT (rpcev->ev_fired == 0);
1297 rpcev->ev_fired = 1;
1298 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1299 -EINTR : ev->status;
1300 swi_schedule_workitem(&crpc->crpc_wi);
1302 cfs_spin_unlock(&crpc->crpc_lock);
1305 case SRPC_REQUEST_RCVD:
1306 sv = rpcev->ev_data;
1308 LASSERT (rpcev == &sv->sv_ev);
1310 cfs_spin_lock(&sv->sv_lock);
1312 LASSERT (ev->unlinked);
1313 LASSERT (ev->type == LNET_EVENT_PUT ||
1314 ev->type == LNET_EVENT_UNLINK);
1315 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1316 sv->sv_shuttingdown);
1318 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1319 buffer->buf_peer = ev->initiator;
1320 buffer->buf_self = ev->target.nid;
1322 sv->sv_nposted_msg--;
1323 LASSERT (sv->sv_nposted_msg >= 0);
1325 if (sv->sv_shuttingdown) {
1326 /* Leave buffer on sv->sv_posted_msgq since
1327 * srpc_finish_service needs to traverse it. */
1328 cfs_spin_unlock(&sv->sv_lock);
1332 cfs_list_del(&buffer->buf_list); /* from sv->sv_posted_msgq */
1333 msg = &buffer->buf_msg;
1334 type = srpc_service2request(sv->sv_id);
1336 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1337 (msg->msg_type != type &&
1338 msg->msg_type != __swab32(type)) ||
1339 (msg->msg_magic != SRPC_MSG_MAGIC &&
1340 msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1341 CERROR ("Dropping RPC (%s) from %s: "
1342 "status %d mlength %d type %u magic %u.\n",
1343 sv->sv_name, libcfs_id2str(ev->initiator),
1344 ev->status, ev->mlength,
1345 msg->msg_type, msg->msg_magic);
1347 /* NB can't call srpc_service_recycle_buffer here since
1348 * it may call LNetM[DE]Attach. The invalid magic tells
1349 * srpc_handle_rpc to drop this RPC */
1353 if (!cfs_list_empty(&sv->sv_free_rpcq)) {
1354 srpc = cfs_list_entry(sv->sv_free_rpcq.next,
1355 srpc_server_rpc_t, srpc_list);
1356 cfs_list_del(&srpc->srpc_list);
1358 srpc_init_server_rpc(srpc, sv, buffer);
1359 cfs_list_add_tail(&srpc->srpc_list,
1360 &sv->sv_active_rpcq);
1361 srpc_schedule_server_rpc(srpc);
1363 cfs_list_add_tail(&buffer->buf_list,
1364 &sv->sv_blocked_msgq);
1367 cfs_spin_unlock(&sv->sv_lock);
1369 cfs_spin_lock(&srpc_data.rpc_glock);
1370 srpc_data.rpc_counters.rpcs_rcvd++;
1371 cfs_spin_unlock(&srpc_data.rpc_glock);
1374 case SRPC_BULK_GET_RPLD:
1375 LASSERT (ev->type == LNET_EVENT_SEND ||
1376 ev->type == LNET_EVENT_REPLY ||
1377 ev->type == LNET_EVENT_UNLINK);
1380 break; /* wait for final event */
1382 case SRPC_BULK_PUT_SENT:
1383 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1384 cfs_spin_lock(&srpc_data.rpc_glock);
1386 if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1387 srpc_data.rpc_counters.bulk_get += ev->mlength;
1389 srpc_data.rpc_counters.bulk_put += ev->mlength;
1391 cfs_spin_unlock(&srpc_data.rpc_glock);
1393 case SRPC_REPLY_SENT:
1394 srpc = rpcev->ev_data;
1395 sv = srpc->srpc_service;
1397 LASSERT (rpcev == &srpc->srpc_ev);
1399 cfs_spin_lock(&sv->sv_lock);
1401 rpcev->ev_fired = 1;
1402 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1403 -EINTR : ev->status;
1404 srpc_schedule_server_rpc(srpc);
1406 cfs_spin_unlock(&sv->sv_lock);
1416 srpc_check_event (int timeout)
1422 rc = LNetEQPoll(&srpc_data.rpc_lnet_eq, 1,
1423 timeout * 1000, &ev, &i);
1424 if (rc == 0) return 0;
1426 LASSERT (rc == -EOVERFLOW || rc == 1);
1428 /* We can't affort to miss any events... */
1429 if (rc == -EOVERFLOW) {
1430 CERROR ("Dropped an event!!!\n");
1434 srpc_lnet_ev_handler(&ev);
1445 memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1446 cfs_spin_lock_init(&srpc_data.rpc_glock);
1448 /* 1 second pause to avoid timestamp reuse */
1449 cfs_pause(cfs_time_seconds(1));
1450 srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1452 srpc_data.rpc_state = SRPC_STATE_NONE;
1455 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1457 if (the_lnet.ln_server_mode_flag)
1458 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1460 rc = LNetNIInit(getpid() | LNET_PID_USERFLAG);
1463 CERROR ("LNetNIInit() has failed: %d\n", rc);
1467 srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1469 LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1471 rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1473 rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &srpc_data.rpc_lnet_eq);
1476 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1480 rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1483 srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1491 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1497 srpc_shutdown (void)
1503 state = srpc_data.rpc_state;
1504 srpc_data.rpc_state = SRPC_STATE_STOPPING;
1509 case SRPC_STATE_RUNNING:
1510 cfs_spin_lock(&srpc_data.rpc_glock);
1512 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1513 srpc_service_t *sv = srpc_data.rpc_services[i];
1515 LASSERTF (sv == NULL,
1516 "service not empty: id %d, name %s\n",
1520 cfs_spin_unlock(&srpc_data.rpc_glock);
1524 case SRPC_STATE_EQ_INIT:
1525 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1527 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1528 LASSERT (rc == 0); /* the EQ should have no user by now */
1530 case SRPC_STATE_NI_INIT: